feat: 支持按算法独立配置置信度阈值

通过 ALGO_CONF_{ALGO_CODE} 环境变量为每个算法设置独立的 conf_threshold，
未配置的算法回退到全局 CONF_THRESHOLD。推理过程零改动，仅后处理过滤阶段
按 bind.algo_code 使用对应阈值。

当前配置：离岗=0.4（降低漏检），入侵=0.5（减少误报）。

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

main.py

@@ -598,14 +598,24 @@ class EdgeInferenceService:
                     self._logger.debug(f"[推理诊断] batch_data shape={batch_data.shape}, outputs={shapes}, 耗时={inference_time_ms:.1f}ms")
 
                 batch_size = len(chunk)
+
+                # 按算法类型获取每个 item 的独立置信度阈值
+                per_item_conf = [
+                    self._settings.inference.get_conf_threshold(item[2].algo_code)
+                    for item in chunk
+                ]
+
                 batch_results = self._postprocessor.batch_process_detections(
                     outputs,
                     batch_size,
-                    conf_threshold=self._settings.inference.conf_threshold
+                    per_item_conf_thresholds=per_item_conf,
                 )
 
                 total_detections = sum(len(r[0]) for r in batch_results)
-                self._logger.debug(f"[推理] batch_size={batch_size}, 总检测数={total_detections}, conf_thresh={self._settings.inference.conf_threshold}")
+                self._logger.debug(
+                    f"[推理] batch_size={batch_size}, 总检测数={total_detections}, "
+                    f"conf_thresholds={per_item_conf}"
+                )
 
                 for idx, (camera_id, roi, bind, frame, _, scale_info) in enumerate(chunk):
                     boxes, scores, class_ids = batch_results[idx]